匯率預測在經濟領域一直都是難解的問題,近幾十年不斷有新的理論問世,但實證結果卻仍 然難以打敗隨機漫步模型。而匯率資料係屬非線性,非定態及高雜訊的資料,使用非線性的 模型進行預測較為合理,故本文選擇了幾個匯率模型,並以 Gated Recurrent Unit(GRU) 神經 網路的方法測試之。 本文建立了一個兩層 GRU 神經網路,使用兩階段及單一方程式兩種預測方式,搭配兩 個匯率決定模型以及一個均衡匯率模型,測試 GRU 神經網路模型的表現是否優於隨機 漫步模型及最小平方法迴歸。實驗結果發現,在使用兩階段預測方式預測匯率走勢時, 各個期數 GRU 神經網路模型能夠表現得比最小平方法迴歸好,但除了五年後匯率的預 測外,其他期數表現都難以穩定勝過隨機漫步。不過在預測漲跌方向的變換上,預測 期間超過兩年時 GRU 神經網路模型都能有效贏過隨機漫步及最小平方法迴歸。在使用 單一方程式預測時可以則發現,無論是預測匯率走勢還是漲跌方向變換,在預測期間超 過半年時,GRU 神經網路模型普遍都優於隨機漫步模型,卻難以穩定勝過最小平方法迴歸。
Exchange rate forecasting has always been a puzzle in the field of Finance and Economics. Numerous models came out during the last few decades, but empirical works found that there were few evidence these models can defeat the random walk model.On the other hand, exchange rate data is typically nonlinear, non-stationary and contains high noise. Thus it is appropriate to use a nonlinear approach to do prediction. In this paper, we build a two layer Gated Recurrent Unit(GRU) recurrent network model, cooper- ate with a two-step and a single-equation method, and test on two exchange decision model and an equilibrium exchange rate model to see whether the deep learning model can outperform the random walk model and OLS regression model. We discover that, with the two-step method, when comparing MSE, deep learning model can outperform OLS in any period. But except when k = 60, deep learning cannot outperform the random walk model. When it comes to predict the direct of change, deep learn- ing model can outperform random walk model and OLS model when k ≥ 6. With single-equation method, both in MSE and direction of change, deep learning model can outperform the random walk model but struggles to defeat OLS model.